A Review of Convolutional Neural Networks in Equipment Wear Particle Recognition

Ferrography is an important method for equipment fault diagnosis, in which ferrography focuses on the analysis of ferrography images, that is, wear particle analysis. Artificial analysis is generally used for wear particle analysis, but due to the complexity of ferrographic images, artificial analysis can not get more objective and unified results. The computer has the objective and stable analysis ability, through the computer intelligence analysis wear particle theory can obtain the objective unified result, therefore the wear particle intelligence recognition aspect research has been paid close attention to by the researcher. Convolutional Neural Network is one of the most popular deep learning algorithms. As a new technology, it has been widely used in the field of image recognition. Its strong self-learning ability and generalization ability have made a breakthrough in image recognition. With the rapid development of convolution neural network, the technology of wear particles in intelligent recognition has made a great breakthrough. Since 2018, Convolution Neural Network has been widely used in wear intelligent recognition, and has achieved better results than the traditional wear intelligent recognition model. 　This article briefly describes the development history of Convolutional Neural Network and wear particle intelligent recognition, which includes a review of the various important models once proposed and corresponding critical time nodes in the period of both Convolutional Neural Networks from origin to present and wear particle intelligent recognition transforming from adopting traditional methods to adopting deep learning methods. Then, the representative literature of Convolution Neural Network applied to wear particle intelligent recognition in recent years is sorted out from two aspects: based on existing network structure such as LeNet-5, inception-v3, AlexNet, Mask R-CNN and self-designed network structure such as FECNN, Small-scale CNN, CDCNN, WP-DRNet, Non-parametric CNN, which summarizes the model structure and characteristics proposed in these work, analyzes and expounds the main recognition principles of each model, the advantages and disadvantages of each network structure, and their data adoption. In addition, the main research directions for intelligent recognition of wear particles in the future are prospected. It is believed that the future research directions should focus on multisource data fusion, defocused image restoration, equipment wear recognition and semi-supervised learning based on online monitoring of wear particles recognition, and briefly introduced the concepts and application examples of these research directions, which provide a certain reference value for the future research and development of intelligent recognition of wear particles. 　To sum up, it affirms the importance of Convolutional Neural Network method in wear particle intelligent recognition, points out that the wear particle recognition model base on this method in the aspects of the datasets still has high labor cost and non-complete objective, and the present situation of the research on the direction of on-line monitoring of wear particle intelligent recognition. Finally, some suggestions are put forward to promote the improvement of wear intelligent recognition level by following the latest technology in the field of image recognition. It has certain significance for the development of wear particles intelligent recognition.